Generative AI in admin and customer service: where to start

"We should put AI into admin" is a sentence that gets said a lot and decides very little. Generative AI is not a department and not a product: it is a specific capability — reading, drafting, summarising and classifying text — that pays off in some tasks and gets in the way in others. The useful decision is not which tool to buy, but which admin and customer service tasks to start with, and what conditions each one needs to actually work. This article walks through the tasks that usually make sense first, and what each demands in return.

Tasks, not products

The usual route starts with the catalogue: see which "AI-powered" tools exist for the department and pick one. The opposite order works better: look at the daily work and find where the repetitive text piles up. Admin and customer service usually have plenty, because they are departments that live off reading, replying, summarising and filing.

It is also worth separating this from classic automation. Anything that follows fixed rules — an invoice from this supplier goes into this folder — is solved with ordinary automation, no AI required. Generative AI earns its keep where the text varies: every email says the same thing differently, every complaint has its own nuance. That is its territory, and it is the one we cover here.

Drafting: removing the blank page

This is the task that pays off soonest. In admin: replies to frequent requests, payment reminders with a tone suited to each case, supplier communications. In customer service: the draft reply an agent reviews, adjusts and sends. The person goes from writing from scratch to correcting — faster, and far less thankless.

The conditions: written criteria, and review before anything goes out. The AI drafts only as well as the rules it has to follow are written down — house tone, what can and cannot be offered, how a complaint is handled. If those criteria only exist in the head of your most experienced person, the first job is not an AI job: it is writing them down. And everything that goes out to a customer gets looked at first; the draft belongs to the machine, the reply belongs to the person.

Summarising: compressing what nobody has time to read

The second family of tasks: the long email thread before replying, a customer's history before a call, meeting minutes, the case file someone has to catch up on. Here the AI works on material you give it, not on its memory, which makes it particularly reliable: the source is right there and checking costs little.

The condition is twofold: the material has to be accessible — a history scattered across three systems does not summarise itself — and the summary gets checked whenever it feeds a decision. A summary for getting oriented can tolerate minor errors; one that decides whether something is accepted or rejected cannot.

Triage and the first reply

The third family faces what comes in: reading each email, request or incident and routing it to the right person, labelled by type and urgency. Today that work is done by a person opening, reading and forwarding — and AI does it well precisely because the incoming text varies.

One step further sits the first reply: confirming receipt, asking for the missing detail, resolving the genuinely simple cases. It is the use that demands the most conditions: clear categories (if two departments fight over the same requests, the AI will inherit the fight), doubtful cases go to a person rather than being forced into a category, and the path to a human always visible — the first reply manages the wait, it does not replace the service. And no improvised content: what the AI does not know for certain, it does not claim.

What stays with people

Just as important as knowing where to start is knowing what stays out of scope: sensitive complaints, exceptions that fit no written rule, decisions with real cost — accepting a return outside the window, negotiating a payment delay — and conversations where the customer needs to feel heard, not processed. In all of them the AI can prepare the ground: pull together the history, suggest a draft. The conversation belongs to the person.

That boundary is not a concession: it is what makes the rest sustainable. A team that sees AI taking the repetitive work and respecting the important work adopts it; a team that suspects it has come to replace them boycotts it quietly.

Five questions before putting generative AI into a process

1. Does the task mostly consist of reading, drafting, summarising or classifying text that varies?
2. Does it repeat often enough to justify setting it up properly?
3. Is the criterion it must follow written down, or does it live in someone's head?
4. Who reviews before anything goes out the door?
5. How will we know it is working: time saved, fewer errors, faster replies?

If the third question fails, that is where the groundwork is — and generative AI applied to muddled criteria produces muddle with better prose.

At Dateliers we prefer to start with one specific task that meets these conditions, measure it and expand from there, rather than roll a tool out across a whole department. If you want to see how we approach it, this is how we work.